What a shared, curated knowledge base changes for an agent — measured, with the payloads shown.
The same query against the same index, with and without the summary layer: eight document fragments the model has to synthesise, or one consolidated summary with a citation per claim, distilled from eleven sources.
A benchmark run that fixed the product it was testing: three defects handed to the agent that could fix them, two merged pull requests back, then re-run and verified — between agents that were never online together.
